SVE 128 bit Implementation of gemm_q4_k_8x8_q8_k kernel - #27491

Open

anubhavfujitsu wants to merge 19 commits into

ggml-org:masterfrom

MonakaResearch:vla_interleaved_kernel_q4_k_q8_k

Commits on Aug 21, 2026

Test Commit - Building and Execution Not verified
anubhavfujitsu
committed
4efdf6c
updated the scale-offset compuation part to SVE, build errors still exist due to bsums array
anubhavfujitsu
committed
03ed7a4
Handled compiler errors until decode_q_Kx8_6bit_scales_sve
anubhavfujitsu
committed
edc8e43
removal of arrays from q8 mul-add pair computation
anubhavfujitsu
committed
71fedfd
SVE code for bias_acc done. Only reorder_acc remains now
anubhavfujitsu
committed
e8495ee
reorder_acc converted to SVE
anubhavfujitsu
committed
c124bcd
datatypes added for reorder_acc
anubhavfujitsu
committed
a6119ad
SVE conversion complete - compilation passes
anubhavfujitsu
committed
88ca4b8
execution is working-commented out the cout
anubhavfujitsu
committed
e0864a7
added return instead of break in switch
anubhavfujitsu
committed
70cfc5a
code added for debug dumps and prints - correct till q8 block loading
anubhavfujitsu
committed
6261739
Finally 128 is running - debug cout commented
anubhavfujitsu
committed
2a2e1d7
Debug log cleanup

NishantPrabhuFujitsu
authored and
anubhavfujitsu
committed
618b65b
review apoints addressed
anubhavfujitsu
committed
939511f
accuracy issue fixed
anubhavfujitsu
committed
60bd01b
resuing q4d_min and q4_d
anubhavfujitsu
committed
9b33d1a
formatting changes SVE 256
anubhavfujitsu
committed
0eebfae
extra new line
anubhavfujitsu
committed
ce43d27
cleanup
anubhavfujitsu
committed
e117e2e